Applied Mathematical Modelling 54 (2018) 752–767

Contents lists available at ScienceDirect

Applied Mathematical Modelling

journal homepage: www.elsevier.com/locate/apm

Very high-order accurate finite volume scheme on curved

boundaries for the two-dimensional steady-state
convection–diffusion equation with Dirichlet condition
Ricardo Costa a,b, Stéphane Clain b,∗, Raphaël Loubère c, Gaspar J. Machado b
a
Institut de Mathématiques de Toulouse, Université Paul Sabatier, 31062 Toulouse, France
b
Centro de Matemática, Universidade do Minho, Campus de Azurém, 4800-058 Guimarães, Portugal
c
CNRS DR researcher, University of Bordeaux Mathematical Institute of Bordeaux (IMB), 33405 Talence, France

a r t i c l e i n f o a b s t r a c t

Article history: Accuracy may be dramatically reduced when the boundary domain is curved and numeri-
Received 23 November 2016 cal schemes require a specific treatment of the boundary condition to preserve the optimal
Revised 10 October 2017
order. In the finite volume context, Ollivier-Gooch and Van Altena (2002) has proposed a
Accepted 11 October 2017
technique to overcome such limitation and restore the very high-order accuracy which
Available online 23 October 2017
consists in specific restrictions considered in the least-squares minimization associated to
Keywords: the polynomial reconstruction. The method suffers from several drawbacks, particularly,
Very high-order finite volume method the use of curved elements that requires sophisticated meshing algorithms. We propose a
Curved boundaries new method where the physical domain and the computational domain are distinct and
Reconstruction for Off-site Data (ROD) we introduce the Reconstruction for Off-site Data (ROD) where polynomial reconstructions
are carried out on the mesh using data localized outside of the computational domain,
namely the Dirichlet condition situated on the physical domain. A series of numerical tests
assess the accuracy, convergence rates, robustness, and efficiency of the new method and
show that the boundary condition is fully integrated in the scheme with a very high-order
accuracy and the optimal convergence rate is achieved.
© 2017 Elsevier Inc. All rights reserved.

1. Introduction

Very high-order finite volume methods require supplemental attention to achieve the optimal order. One of the major
difficulties is the boundary treatment when dealing with curved boundary domains, since polygonal meshes do not exactly
fit the physical domain. Without special attention we observe a dramatic reduction of the accuracy and the method turns out
to be a second-order accurate one [1,2]. Reaching the nominal convergence order of very high-order methods then requires
additional efforts and is of paramount importance nowadays [3,4]. Several critical issues motivate the use of very high-order
approximations with curved boundaries. For the Euler system, it is difficult to compute asymptotic solutions when using
piecewise linear approximations of the geometries [5] even for very fine meshes. Moreover, non-physical approximations
may be obtained when curved boundaries are substituted with piecewise linear straight lines [1,2].

∗
Corresponding author.
E-mail addresses: rcosta@math.univ-toulouse.fr (R. Costa), clain@math.uminho.pt (S. Clain), raphael.loubere@u-bordeaux.fr (R. Loubère),
gjm@math.uminho.pt (G.J. Machado).

https://doi.org/10.1016/j.apm.2017.10.016
0307-904X/© 2017 Elsevier Inc. All rights reserved.
 R. Costa et al. / Applied Mathematical Modelling 54 (2018) 752–767 753

Several technologies to recover the optimal order have then been proposed and extensively tested. Deriving from the
Finite Element approach [6–9], the Discontinuous Galerkin method [10,11] handles curved boundaries with isoparametric
elements first introduced by Bassi and Rebay [1]. A similar approach has also been developed for the Spectral Volume
method [4,12]. In short, the technology is based on two major ingredients: the mesh considers curved elements such that
the boundary of the computational domain fits with the physical boundary (at least up to a given error O(hk )) and the
introduction of nonlinear transformations to map the curved elements to the reference one.
The method is efficient and provides optimal order of convergence but suffers from several drawbacks. The element map-
ping introduces the Jacobian transformations in the volume and interface integrals that are evaluated in the local basis (the
reference element coordinates). Such expressions become cumbersome when dealing with high polynomial degrees and lead
to an additional computational effort. Moreover a change of sign of the Jacobian mapping, i.e. folded or tangled elements,
may occur and disqualifies the transformation [13,14]. The second and more critical drawback is the meshing procedure with
curved elements, that reveals impractical for complex geometries especially for three-dimension configurations [2]. Curved
mesh generation is a today’s challenge [3] and is far from being completely solved [15]. Meshing very complex geometries
with unstructured hybrid grids has still not reached the level of commercial grid generators.
Alternative methods have been then proposed to avoid the nonlinear mapping and the curved mesh generation.
• In [2], the authors use the computational polygonal domain in place of the physical domain but modify the normal
vector involved in the wall boundary condition (see also [16]). Accuracy improvement is obtained but, unfortunately, the
method has only been tested with quadratic boundaries and seems to be, at most, a third-order approximation since it
considers local curvature approximations, i.e. second derivatives in the Taylor expansion.
• Another promising method is the so-called “Extensions from Subdomains” introduced in [17–19]. The idea is to derive
a new Dirichlet condition on the computational domain edge from the one evaluated on the physical boundary. An
additional contribution is obtained from the integration of the solution over a path linking points xcomp and xphys lying
on their respective borders. The main advantage is that no local mapping or curved element is required but an extension
of the numerical approximation has to be evaluated in order to perform the integration of the approximation gradient
outside the computational domain. Unfortunately, the method is only available for second-order operators with diffusion
or viscosity term. The second drawback is the necessity to define path families between the edges of the computational
mesh boundary and the physical one. Such procedure introduces constrained local minimization operations to define
local one-to-one mappings x → a(x) from the boundary edges and the associated pieces of the physical boundary. Then,
the paths are derived from linear interpolation [x, a(x)]. At last, numerical integrations over the paths are required to
compute the additional contributions to update the associate Dirichlet condition on the computational boundary.
• We also mention the unfitted Finite Element method where an inner interface corresponding to a change of material
properties does not fit with the mesh [20]. Two linear reconstructions are considered in the same cell where the interface
is localized (one for each side). Then a specific weak formulation is derived taking into account the inner interface while
flux conservation and continuity is assume to solve the problem. Extension for higer order finite elements is considered
in [21] where a local transformation, very similar to the isoparametric one, is considered to fit a local discretized interface
with the physical one using the isoparametric transformation proposed in [22]. The main drawback of the method is the
use of isoparametric transformation that involves more complex algebra and additional computational costs.
• In the context of the finite volume method involving k-exact polynomial reconstructions, the pioneer work of Ollivier-
Gooch and Van Altena [23] gives rise to a very high-order finite volume method dealing with curved boundary for the
convection–diffusion equation and Euler and Navier–Stokes systems [24–26]. The method does not require any geomet-
rical transformation but the mesh has to be composed with curved cells that fits the physical boundary. Likewise for the
Discontinuous Galerkin method, it represents a severe drawback due to the difficulties to provide such curved elements
for complex geometries. Moreover, the method suffers from other problems. Indeed, one has to perform numerical inte-
grations over the curved elements to evaluate any source term or the initial condition mean-values and on the curved
boundaries to calculate the numerical fluxes. Integration on a piece of curves boundary for the two-dimensional situa-
tion requires an extra-effort for localizing the Gauss points, but the problem turns out to be cumbersome when dealing
with a three-dimensional domain [27]. Integration over the cell is a major difficulty. Indeed, while Gaussian points are
well-located for simple geometries, it seems almost impossible to derive numerical quadratures rules for generic curved
element except for special situations such as pieces of circles or spheres [25].
We propose a new and simple treatment of Dirichlet conditions in the context of very high-order finite volume methods
with a curved boundary domain. Because this paper is a proof of concept, we choose the simplest situation, namely the
steady-state convection–diffusion with Dirichlet condition on two-dimensional curved domains. Extensions such as three-
dimensional geometries, Neumann or mixed condition, and non-stationary systems will be considered in the future.
As in [17,27], we consider two distinct regions: the physical domain, where the continuous problem takes place, and the
computational polygonal domain, where the discretization is designed and the numerical solution is evaluated. Obviously,
Dirichlet conditions prescribed on the physical boundary have to be transferred in some way to the computational domain.
The corner-stone of our work is the design of a specific polynomial reconstruction that takes the real boundary condition
into account. Therefore we perform a reconstruction with data that are not all localized on the mesh and name the method
“Reconstruction for Off-site Data” with acronym “ROD” to highlight that data are not supported by the computational do-
main.
754 R. Costa et al. / Applied Mathematical Modelling 54 (2018) 752–767

Fig. 1. Mesh notation with edge and cell reference points (blue dots), Gauss points (red dots associated to a two-point quadrature rule), and unit normal
vectors (dashed lines). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article).

In some way, isoparametric element method involves two polynomials: one for the curved boundary and another for
the solution approximation. Conversely, the method proposed in this work uses a unique polynomial for the two associ-
ated operations. Despite, the study in [27] presents some fundamental ingredients for the polynomial reconstructions under
curved boundaries, the present paper proposes a simpler approach where all the integration procedures are performed on
the polygonal domain.
This document is divided in seven sections. After the introduction, we formulate in Section 2 the problem we will con-
sider and the mesh notations, while we present the polynomial reconstructions in Section 3. Section 4 is dedicated to the
curved boundary where we detail the inclusion of the Dirichlet condition prescribed on the physical boundary. We present
in Section 5 the very high-order finite volume and in Section 6 the numerical experiments. We end the document with the
conclusions of this work and some perspectives.

2. Problem formulation and geometry

Let  be an open bounded domain of R2 with boundary ∂ . We assume that the boundary is a regular Jordan curve
which admits local parametrizations. We seek function φ ≡ φ (x), x ≡ (x1 , x2 ), solution of the steady-state convection–diffusion
equation

∇ · (uφ − κ∇φ ) = f, in , (1)

where u = (u1 , u2 ) ≡ (u1 (x ), u2 (x ) ) is a velocity field, κ ≡ κ (x) is a diffusion coefficient, and f ≡ f(x) is a regular source term
in , and where a Dirichlet condition is prescribed on the boundary ∂  with a given regular function φD ≡ φD (x ).
A mesh M is a set of I non-overlapping convex polygonal cells ci , without gap, i ∈ CM = {1, . . . , I}, and we denote by

 = ci
i∈C M

the computational domain and by ∂  the computational boundary associated to the mesh M.  should be a represen-
tative domain approximation of  and ∂  is the associated approximation of ∂ . To this end, we assume that the nodes
on ∂  also belong to ∂ . We adopt the notations detailed hereafter (see Fig. 1):

• for any cell ci we denote by ∂ ci its boundary and by |ci | its area; the reference cell point is denoted by mi which can be
any point in ci (in the present work we consider the centroid);
• two different cells ci and cj share a common edge eij whose length is denoted by |eij | and ni j = (n1,i j , n2,i j ) is the unit
normal vector to eij outward to ci , i.e. ni j = −n ji ; the reference edge point is mij which can be any point on eij (in the
present work we consider the midpoint); if an edge of ci belongs to the boundary ∂  , the index j is replaced by letter
D;
• for any edge eij , we denote by qij,r , r = 1, . . . , R, the Gaussian integration points and ζ r the associated weights;
• for any cell ci we associate the index set of neighbor cells ν (i) ⊂ {1, , I} ∪ {D} such that j ∈ ν (i) if eij is a common edge
between cells ci and cj or with the boundary ∂  if j = D.

We enhance that  is not a polygonal domain. So, the physical domain  and its polygonal approximation  do not
coincide, and this usually leads to a significant accuracy degradation of the numerical approximation.

3. Polynomial reconstruction

The polynomial reconstruction is a powerful tool to provide an accurate local representation of the underlying solu-
tion, see [28,29] for unstructured grids and hyperbolic problems. In [30] a methodology was proposed in the context of
convection–diffusion problems to achieve very high accurate approximations of the gradient fluxes and take into account
 R. Costa et al. / Applied Mathematical Modelling 54 (2018) 752–767 755

boundary conditions. The authors introduced different types of polynomial reconstructions namely the conservative recon-
struction in cells and on boundary edges and the non-conservative reconstruction on inner edges, in order to compute ap-
proximations of the convective and the diffusive fluxes. In this work we mainly follows this methodology of reconstruction
but applied in the specific but important case of curved boundaries.

3.1. Stencil and data

A stencil is a collection of cells situated in the vicinity of a reference geometrical entity, for instance an edge or a
cell, where the number of elements of the stencil shall depend on the degree d of the polynomial function we intend to
construct. So, for each edge eij and cell ci we associate stencils Sij and Si , respectively, consisting of the indices of neighbor
cells.

Remark 3.1. A polynomial reconstruction of degree d requires nd = (d + 1 )(d + 2 )/2 coefficients in 2D. So, in practice, a
stencil consists of the Nd closest cells to each geometrical entity (edge or cell) with Nd ≥ nd (we consider Nd ≈ 1.5nd for the
sake of robustness).

To compute the polynomial reconstructions we need the data associated to each cell of the stencil. To this end, we
assume that vector = (φi )i∈CM gathers the approximation of the mean-value of φ over each cell, i.e.

1
φi ≈ φ dx.
|ci | ci

3.2. Conservative reconstruction for cells

For each cell ci , the local d-th degree polynomial approximation of the underlying solution φ , based on vector , is
defined as
  
φi ( x ) = φ i + Rαi (x − mi )α − Miα ,
1≤|α|≤d
α α
where α = (α1 , α2 ) with |α| = α1 + α2 and the convention xα = x1 1 x2 2 . Vector Ri = (Rα )
i 1≤|α|≤d
gathers the polynomial

coefficients, and Miα = |c1 | c (x − mi )α dx in order to guarantee the conservation property
i i

1
φi (x ) dx = φi .
|ci | ci

For a given stencil Si , we consider the quadratic functional

  2
 1
Ei ( Ri ) = φi ( x ) d x − φ q . (2)
q∈Si
|cq | cq

We denote by R
(x ) the polynomial which

i the unique vector which minimizes the quadratic functional (2) and we set φi
corresponds to the best approximation in the least squares sense.

3.3. Non-conservative reconstruction for inner edges

For each inner edge eij , the local d-th degree polynomial approximation of the underlying solution φ , based on vector ,
is defined as

φi j ( x ) = Rαi j (x − mi j )α ,
0≤|α|≤d

where vector Ri j = (Rα )

i j 0≤|α|≤d
gathers the polynomial coefficients (notice that in this case |α | starts with 0). For a given
stencil Sij with #Si j elements and vector ωi j = (ωi j,q )q=1,...,#Si j of positive weights of the reconstruction, we consider the
quadratic functional
  2
 1
Ei j ( Ri j ) = ωi j,q φi j (x ) dx − φq . (3)
q∈Si j
|cq | cq

We denote by R  (x ) the polynomial which

i j the unique vector which minimizes the quadratic functional (3) and we set φij
corresponds to the best approximation in the least squares sense.

Remark 3.2. The weights are fundamental to provide the maximum principle and the stability of the scheme. In the simu-
lations, we shall prescribe ωi j,q = 3 for the adjacent cells while we set ωi j,q = 1 for the other cells of the stencil. A deeper
analysis to justifying the weights choice is given in [27,30].
756 R. Costa et al. / Applied Mathematical Modelling 54 (2018) 752–767

Fig. 2. The physical boundary ∂  and the boundary edge eiD with Gauss points (red dots associated to a two-point quadrature rule) and edge midpoints
miD (blue dots): left – situation when curved cells are considered; right – situation considered in this work with straight-edge cell and an associated
collocation point biD associated to miD . (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this
article).

3.4. Conservative reconstruction for computational boundary edges

We treat the computational boundary edges in a particular way in order to take into account the prescribed Dirichlet
condition. For each boundary edge eiD on ∂  , the local d-th degree polynomial approximation of the underlying solution
φ is defined as
  
φ iD ( x; ψiD ) = ψiD + RαiD (x − miD )α − MiαD ,
1≤|α|≤d

where vector RiD = (Rα )

iD 1≤|α|≤d
gathers the polynomial coefficients, ψiD ∈ R is a free parameter which shall be set later,
and MiαD = ( piD − miD )α in order to guarantee the conservation property
ψiD = φ iD ( piD ; ψiD ), (4)
with piD a given collocation point. The crucial point is that piD will be a distinct point from midpoint miD ∈ eiD . For a given
stencil SiD with #SiD elements and vector ωiD = (ωiD,q )q=1,...,#SiD of positive weights of the reconstruction, we consider the
quadratic functional
  2
 1
EiD ( RiD ) = ωiD,q φ i D ( x ; ψ i D ) d x − φq . (5)
q∈SiD
|cq | cq

We denote by R
iD the unique vector which minimizes the quadratic functional (5) and we set
φiD (x; ψiD ) the polynomial
which corresponds to the best approximation in the least squares sense for the given parameter ψ iD and point piD .

Remark 3.3. The motivation for introducing the weights either for the case of a non-conservative polynomial reconstruc-
tion or for a conservative polynomial reconstruction for computational boundary edges is presented in [30] as well as the
importance to set larger values for the adjacent cells. We refer the reader to [30] for more details.

4. Reconstructions for Off-site Data (ROD)

4.1. The Ollivier-Gooch and Van Altena method

An accurate approximation of boundary conditions on curved boundaries is of paramount importance when dealing with
very high-order methods, since to approximate a curved boundary with a polygonal mesh generally leads to a second-order
approximation [3]. The seminal paper of Ollivier-Gooch and Van Altena [23] introduces a technique for constraining the
least-squares problem associated to the polynomial reconstructions on the boundary elements. Such approach requires that
the reconstructed solution satisfies the boundary condition exactly at the flux integration points (collocation points) [25]. In
order to properly represent the boundary condition, the mesh has to fit with the physical boundary, that is, the edges of the
mesh are curved for matching the real boundary (see Fig. 2). As mentioned in the introduction, this approach brings several
drawbacks: one has to carefully design Gauss quadrature procedures to take into account curved boundaries for preserving
the accuracy [26], and designing boundary-fitted mesh for non polygonal domains is still nowadays a difficult task [3]. In
this work we avoid the difficult construction of boundary-fitted mesh and solely work on the easy to construct polygonal
mesh.

4.2. The ROD method

We propose a different technique to prescribe Dirichlet conditions on curved boundaries. The main idea is to distinguish
the computational domain from the physical one, and, perform all computations on the polygonal cells, but taking into
account the information located on the physical boundary via the polynomial reconstructions. “Off-site Data” method is
 R. Costa et al. / Applied Mathematical Modelling 54 (2018) 752–767 757

meant to remind that the scheme and the solutions are acting on the computational domain  , but including information
which is not associated to any geometrical entity of  (cell, edge, or point). The main advantages are:
• numerical integration of flux or functions are only carried out on the polygonal domain and not on the complex physical
domain;
• no curved elements are required, only the computational polygonal mesh is necessary;
• no geometrical transformations are required involving possibly complex Jacobian functions for the integrals;
• no Gaussian points on the physical boundary ∂  are required;
• the method design does not depend on the number of spacial dimensions.
The technique is intrinsically associated to the conservative polynomial reconstruction given in Section 3.4 where, for
a given edge eiD of the computational boundary ∂  , we define the polynomial approximation
φiD (x; ψiD ) depending on
parameter ψ iD and point piD which satisfies the conservation property (4). We also mention that biD stands for a point on
the physical boundary ∂ , somewhere facing edge eiD as depicted in Fig. 2-right while miD stands for the edge midpoint.
The Dirichlet condition will be enforced via a clever choice of the values of free parameter ψ iD and point piD . As a conse-
quence one has to design a procedure to compute the free parameter such that we simultaneously satisfy the conservation
and provide a very high-order approximation of the boundary condition.

4.2.1. Second-order approximation

A simple approach consists in using piD = miD and in setting the free parameter as ψiD = φD (miD ). Such a choice pro-
vides no more than a second-order convergence rate since miD does not represent exactly the physical boundary. Moreover,
we need an extension of function φD in the neighborhood of the boundary to guarantee that φD (miD ) makes sense since
miD ∈ ∂ .

4.2.2. Very high-order approximation: ROD1

In order to enforce the Dirichlet condition more accurately, we now set piD = biD and the free parameter ψiD = φD (biD ).
Notice that the flux integration will nonetheless be computed on the straight edge of the computational mesh as presented
in Section 5.2 and not on the physical boundary as in [23]. We expect a very high-order accuracy since the reconstruction
satisfies the Dirichlet condition at one point associated to the true physical boundary.

4.2.3. Very high-order approximation: ROD2

The major drawback of the method ROD1 is that the least-squares problem (5) is based on point piD that depends on
the physical domain boundary and if the physical boundary evolves, for instance for time dependent moving domains or
interface tracking problems, one has to rebuild the whole reconstruction procedure for boundary cells/edges. We improve
the previous technique by decoupling the Dirichlet condition from the interpolation problem still preserving the very high-
order accuracy. We start again by setting the collocation point piD = miD as in the second-order method. Hence the least-
squares procedure (5) no longer depends on the physical boundary position. Next the free parameter ψ iD is computed in a
special way. To this end, let us introduce the functional
ψiD → B iD ( biD ; ψiD ) =

φ i D ( b i D ; ψ i D ) − φD ( b i D ) . (6)
Notice that BiD is affine with respect to ψ iD .
We now seek for ψi∗D as the unique solution which satisfies the affine problem
B iD ( biD ; ψiD ) = 0. (7)
We compute the solution of this problem by taking two values ψi1D , ψi2D ∈ R, ψi1D = ψi2D , and after some algebraic simplifi-
cations one get
ψi2D BiD (biD ; ψi1D ) − ψi1D BiD (biD ; ψi2D )
ψi∗D = . (8)
BiD (biD ; ψi1D ) − BiD (biD ; ψi2D )
In other words, we adjust the free parameter at point miD to satisfy the Dirichlet condition at biD . In practice, we take
ψi1D = φD (miD ) and ψi2D = φD (miD ) + 1 since ψi∗D is supposed to be close to φD (miD ).
Remark 4.1. The methods ROD1 and ROD2 mainly differ in the structure of the matrix that compute the polynomial coeffi-
cients with respect to the data and the Dirichlet condition. In the first case, the matrix depends on the position of the point
where the Dirichlet is evaluated but does not require the additional treatment given in relation (8). On the contrary, the
ROD2 method provides a matrix that does not depend of the position of the Dirichlet condition but an extra-computational
effort is necessary to fix the free parameter with (8). To sum-up, if one considers a fix curved domain, the ROD1 method
is more efficient whereas the ROD2 technique is well-adapted to situations where the physical boundary changes with time
or during an iterative process.

Remark 4.2. The coefficients R

iD of the polynomial function
φiD (x; ψiD ) are obtained as the matrix-vector product between
the Moore-Penrose matrix associated to the least-square problem and the vector of cell values in the stencil [30]. The matrix
structure does not depend on the physical boundary position by construction but only depends on the computational mesh.
The Dirichlet condition is only prescribed via functional (6) and satisfies condition (7).
758 R. Costa et al. / Applied Mathematical Modelling 54 (2018) 752–767

5. Very high-order finite volume scheme

5.1. Generic finite volume scheme

To obtain a finite volume scheme, Eq. (1) is integrated over each cell ci and applying the divergence theorem we get
 
(uφ − κ∇φ ) · ni ds = f dx,
∂ ci ci

where ∂ ci is the cell boundary and ni is the associated outward unit vector. Considering the Gaussian quadrature with
R ∈ N∗ points for the line integrals, i.e. of order 2R, we get the residual expression
 
 
R 
|ei j | ζ C
r Fi j,r + FD
i j,r − fi |ci | = O (h2i R ), (9)
j∈ν (i ) r=1

with the physical fluxes given by

FCi j,r = u(qi j,r ) · ni j φ (qi j,r ) and i j,r = −κ (qi j,r )∇φ (qi j,r ) · ni j ,
FD
and with hi = max j∈ν (i ) |ei j |, while fi stands for an approximation of order 2R of the mean value of f over cell ci . Notice
that if cell ci is not triangular, we split it into sub-triangles which share the cell centroid as a common vertex and apply
the quadrature rule on each sub-triangle as in [31]. Using the different polynomial reconstructions see previous sections, we
design the numerical scheme with two main ingredients: the flux computation and the solver. We use a similar technique
proposed in [27,30], particularly the matrix-free approach is adopted, based on the residual operator construction.

5.2. Numerical fluxes

Numerical fluxes are computed with respect to the edges:

• for the inner edges eij , the fluxes at the quadrature point qij,r write
 +   − 
FiCj,r = u(qi j,r ) · ni j φi (qi j,r ) + u(qi j,r ) · ni j φ j (qi j,r ),

 (q ) · n ;
FiDj,r = −κ (qi j,r )∇ φ ij i j,r ij

• for the boundary edges eiD , the fluxes at the quadrature point qiD,r write
+ −
FiCD,r = [u(qiD,r ) · niD ] φ i (qiD,r ) + [u (qiD,r ) · niD ] φ iD (qiD,r ),
 (q ) · n .
FiDD,r = −κ (qiD,r )∇ φ iD iD,r iD

Notice that all the fluxes are computed on the edges of the computational domain without any reference to the physical
 .
domain. The Dirichlet condition on ∂  is implicitly contained in the polynomial reconstructed function φ iD

5.3. Residual operator and solver

For any vector in RI , we define the residual operators for cells ci , i = 1, . . . , I, as

 
 
R 
Gi ( ) = |ei j | ζ C
r Fi j,r + FiDj,r − f i |ci |,
j∈ν (i ) r=1

which corresponds to the finite volume scheme (9) cast in residual form. Gathering all the components of the residuals
provides a global affine operator G ( ) = (Gi ( ) )i∈CM and we seek vector ∈ RI , solution of the problem G ( ) = 0. The
GMRES method, powered by a preconditioning matrix, is carried out to compute an approximation of as in [27,30].

6. Numerical results

In order to validate the implementation of the methods and assess the accuracy and the convergence rates, we manu-
facture several analytic solutions on specific domains which require the computation of an associated source term to satisfy
Eq. (1). Vector = (φi )i∈CM gathers the numerical approximations of the mean values of φ while vector = (φ i )i∈CM

gathers the exact mean values φ i of φ , that is φ i = (1/|ci | ) c φ dx.
i
The normalized L1 - and L∞ -norm errors, denoted by E1 and E∞ , are computed, respectively, as

i∈C M |φi − φ i ||ci | maxi∈CM |φi − φ i |
E1 ( M ) =  and E∞ ( M ) =  .
i∈CM |φ i ||ci | i∈CM |φ i ||ci |
 R. Costa et al. / Applied Mathematical Modelling 54 (2018) 752–767 759

Fig. 3. Manufactured solution (left panel) and source term (right panel) for the low Péclet number test case. (For interpretation of the colormap in this
figure, the reader is referred to the web version of this article).

Fig. 4. Coarse uniform triangular Delaunay mesh prescribed for the annulus domain.

The convergence rate for the normalized L1 - and L∞ -norm errors between two different meshes M1 and M2 , with DOF1
and DOF2 degrees of freedom, respectively, where DOF1 = DOF2 , is evaluated as
| log(Eα (M1 )/Eα (M2 ))|
Oα ( M1 , M2 ) = 2 , α ∈ {1, ∞}.
| log(DOF1 /DOF2 )|
In all the simulations, the weights in functionals (3) and (5) are set to ωi j,q = 3, i ∈ CM , j ∈ ν (i), q ∈ Sij , if eij is an edge of cq
and ωi j,q = 1, otherwise.
For the sake of simplicity, we name the method given in Section 4.2.1 as “second-order method”, the one given in
Section 4.2.2 as “ROD1 method”, and the proposed method, presented in Section 4.2.3, as “ROD2 method”. The method-
ology of validation consists in observing the rates of convergence under mesh refinement when very high-order polynomial
reconstructions (Pk , k = 1, 3, 5) are employed for different strategies dealing with Dirichlet boundary conditions defined on
760 R. Costa et al. / Applied Mathematical Modelling 54 (2018) 752–767

Table 1
Annulus problem for low Péclet number – Errors and convergence rates for the second-order, ROD1, and ROD2 methods with uniform trian-
gular Delaunay meshes.

Second-order method (Section 4.2.1)

P1 P3 P5

DOF E1 O1 E∞ O∞ E1 O1 E∞ O∞ E1 O1 E∞ O∞

736 2.17E−02 – 4.56E−02 – 3.87E−03 – 9.88E−03 – 3.84E−03 – 9.82E−03 –

2828 6.62E−03 1.77 1.24E−02 1.94 1.04E−03 1.95 2.79E−03 1.88 1.04E−03 1.94 2.81E−03 1.86
11,500 1.52E−03 2.10 3.59E−03 1.76 2.63E−04 1.97 7.11E−04 1.95 2.63E−04 1.97 7.12E−04 1.96
45,248 4.07E−04 1.93 1.01E−03 1.84 6.74E−05 1.99 1.85E−04 1.97 6.74E−05 1.99 1.85E−04 1.97
177,880 1.35E−04 1.61 3.08E−04 1.74 1.71E−05 2.01 4.71E−05 2.00 1.71E−05 2.01 4.71E−05 2.00

Reconstruction for Off-site Data (ROD1) method (Section 4.2.2)

P1 P3 P5

DOF E1 O1 E∞ O∞ E1 O1 E∞ O∞ E1 O1 E∞ O∞

736 2.06E−02 – 4.55E−02 – 1.55E−04 – 9.14E−04 – 5.06E−05 – 2.22E−04 –

2828 6.21E−03 1.78 1.21E−02 1.97 9.63E−06 4.13 6.51E−05 3.93 3.30E−07 7.48 4.95E−06 5.65
11,500 1.43E−03 2.10 3.49E−03 1.77 4.92E−07 4.24 3.93E−06 4.00 5.43E−09 5.85 9.15E−08 5.69
45,248 3.82E−04 1.93 1.02E−03 1.80 3.72E−08 3.77 4.13E−07 3.29 9.38E−11 5.93 2.84E−09 5.07
177,880 1.28E−04 1.59 3.27E−04 1.66 3.76E−09 3.35 2.44E−08 4.14 1.76E−12 5.81 4.21E−11 6.15

Reconstruction for Off-site Data (ROD2) method (Section 4.2.3)

P1 P3 P5

DOF E1 O1 E∞ O∞ E1 O1 E∞ O∞ E1 O1 E∞ O∞

736 2.07E−02 – 4.56E−02 – 1.54E−04 – 9.01E−04 – 4.74E−05 – 2.13E−04 –

2828 6.22E−03 1.78 1.21E−02 1.97 9.67E−06 4.11 6.47E−05 3.91 3.25E−07 7.40 4.91E−06 5.60
11,500 1.43E−03 2.10 3.50E−03 1.77 4.92E−07 4.25 3.92E−06 4.00 5.39E−09 5.84 9.17E−08 5.68
45,248 3.82E−04 1.93 1.02E−03 1.80 3.71E−08 3.77 4.13E−07 3.29 9.43E−11 5.91 2.84E−09 5.07
177,880 1.28E−04 1.59 3.27E−04 1.66 3.76E−09 3.35 2.44E−08 4.13 1.76E−12 5.82 4.21E−11 6.15

the physical curved domains. Only smooth solutions of the steady-state two-dimensional convection–diffusion equation are
considered. The different curved physical domains are
• An annulus domain in Section 6.1. We simulate with triangular (possibly refined) grids low Péclet number, high Péclet
number and pure convection problems.
• A rose-shaped domain in Section 6.2. A deformation of the annulus allows to define branches on the interior and exterior
boundaries. We simulate with triangular and quadrilateral grids two different cases, a rose with three-three branches and
another five-three branches.

6.1. Annulus domain

In this first set of numerical tests we consider an annulus domain with center at (0, 0) characterized by the interior and
exterior circumferences I and E , respectively, with radius rI = 0.5 and rE = 1. For the convection–diffusion problem (1),
we prescribe a constant radial velocity u and κ = 1. We then seek for a manufactured solution, invariant by rotation, given
be
 (2r − (rE + rI ))
φ (x1 , x2 ) = a exp(ur ) + exp(−ur ) + b , r  ≡ r  (r ) = ,
( rE − rI )
with r 2 = x21 + x22 such that r  ∈ [−1, 1]. We also prescribe homogeneous Dirichlet conditions on the two boundaries I and
E and deduce that b = − exp(u ) − exp(−u ) while a = 1/(exp(u ) + exp(−u ) − 2 ) guarantees the property φ ∈ [−1, 0] in .
The associated source term f is obtained after substituting the solution into Eq. (1).

Low Péclet number. We first address the low Péclet number situation setting u = 1. We plot in Fig. 3 the manufactured
solution and the source term. The simulations were carried out with successive refined uniform triangular Delaunay meshes
(see Fig. 4). Observe that boundary vertices belong to the true physical domain boundary.
We report in Table 1 the errors and the convergence rates for the second-order, and the two ROD methods. The second-
order approach provides at most a second-order convergence for both error norms, whatever the degree of the polynomial
reconstruction. These results are expected since the Dirichlet condition is affected with a mismatch of order O (h2 ) due to
the erroneous location with respect to the physical boundary. The two other methods recover the optimal order and achieve
an effective second-, fourth-, and sixth-order convergence rates for P1 , P3 , and P5 polynomial reconstructions, respectively,
while no oscillations are reported. The accuracy of both methods are quite comparable and clearly overcome the second-
order limitation expected when dealing with the curved boundary with non-fitted cells.
 R. Costa et al. / Applied Mathematical Modelling 54 (2018) 752–767 761

Fig. 5. Manufactured solution (left panel) and source term (right panel) for the high Péclet number test case. (For interpretation of the colormap in this
figure, the reader is referred to the web version of this article).

Fig. 6. Coarse non-uniform triangular Delaunay mesh prescribed for the annulus domain for the high Péclet number test case.

High Péclet number. Large Péclet number is prescribed taking u = 10 and we plot in Fig. 5 the manufactured solution and the
source term. This test addresses the scheme robustness and accuracy to preserve the boundary condition when dealing with
small boundary layers with respect to the dimension of the whole geometry. The simulations were carried out with suc-
cessive refined Delaunay meshes plotted in Fig. 6 where, again, the boundary vertices belong to the physical boundary.The
meshes are refined close to the boundaries to better capture the boundary layers.
In Table 2, we report the errors and the convergence rates for the three methods. As for the low Péclet problem the
second-order method reaches at most a second-order convergence for both error norms while the two ROD methods achieve
an effective second-, fourth-, and sixth-order convergence rates for P1 , P3 , and P5 polynomial reconstructions, respectively.
We conclude that the very high-order methods effectively handle large Péclet number situations achieving optimal conver-
gence rates without any oscillation.
762 R. Costa et al. / Applied Mathematical Modelling 54 (2018) 752–767

Table 2
Annulus problem for high Péclet number — Errors and convergence rates for the second-order, ROD1, and ROD2 methods with adapted
triangular Delaunay meshes.

Second-order method (Section 4.2.1)

P1 P3 P5

DOF E1 O1 E∞ O∞ E1 O1 E∞ O∞ E1 O1 E∞ O∞

4292 1.16E−02 – 3.63E−02 – 6.36E−04 – 4.04E−03 – 1.20E−03 – 2.07E−03 –

16,398 2.73E−03 2.16 1.32E−02 1.51 2.11E−04 1.64 6.15E−04 2.81 2.48E−04 2.36 4.67E−04 2.22
63,364 3.57E−04 3.01 4.30E−03 1.66 5.74E−05 1.93 1.28E−04 2.32 6.10E−05 2.08 1.16E−04 2.06
250,732 1.17E−04 1.62 1.20E−03 1.86 1.53E−05 1.93 3.00E−05 2.11 1.53E−05 2.01 2.93E−05 2.01
996,002 8.10E−05 0.53 3.07E−04 1.98 3.85E−06 2.00 7.38E−06 2.03 3.85E−06 2.01 7.35E−06 2.00

Reconstruction for Off-site Data (ROD2) method (Section 4.2.2)

P1 P3 P5

DOF E1 O1 E∞ O∞ E1 O1 E∞ O∞ E1 O1 E∞ O∞

4292 1.08E−02 – 3.46E−02 – 9.47E−04 – 2.43E−03 – 2.93E−04 – 5.83E−04 –

16,398 2.52E−03 2.17 1.27E−02 1.49 5.61E−05 4.22 2.14E−04 3.63 5.88E−06 5.83 1.48E−05 5.48
63,364 3.17E−04 3.07 4.19E−03 1.65 4.92E−06 3.60 1.68E−05 3.77 8.08E−08 6.34 2.69E−07 5.93
250,732 1.25E−04 1.35 1.17E−03 1.85 1.23E−07 5.36 1.42E−06 3.60 1.10E−09 6.24 5.65E−09 5.62
996,002 8.39E−05 0.58 3.01E−04 1.97 7.15E−09 4.13 8.87E−08 4.02 1.00E−11 6.82 8.52E−11 6.08

Reconstruction for Off-site Data (ROD2) method (Section 4.2.3)

P1 P3 P5

DOF E1 O1 E∞ O∞ E1 O1 E∞ O∞ E1 O1 E∞ O∞

P1 P3 P5

DOF E1 O1 E∞ O∞ E1 O1 E∞ O∞ E1 O1 E∞ O∞

4292 1.08E−02 – 3.46E−02 – 9.47E−04 – 2.43E−03 – 2.92E−04 – 5.83E−04 –

16,398 2.52E−03 2.17 1.27E−02 1.49 5.61E−05 4.22 2.14E−04 3.63 5.88E−06 5.83 1.48E−05 5.48
63,364 3.17E−04 3.07 4.19E−03 1.65 4.92E−06 3.60 1.68E−05 3.77 8.08E−08 6.34 2.69E−07 5.93
250,732 1.25E−04 1.35 1.17E−03 1.85 1.23E−07 5.36 1.42E−06 3.60 1.10E−09 6.24 5.65E−09 5.62
996,002 8.39E−05 0.58 3.01E−04 1.97 7.15E−09 4.13 8.87E−08 4.02 1.00E−11 6.82 8.52E−11 6.08

Fig. 7. Manufactured solution (left panel) and source term (right panel) for the pure convection test case. (For interpretation of the colormap in this figure,
the reader is referred to the web version of this article).

Pure convection. We address the pure convection situation setting κ = 0 and u = 1 and plot in Fig. 7 the manufactured
solution and the source term. The simulations were carried out with successive refined uniform triangular Delaunay meshes
presented in Fig. 4 and, as in the previous tests, the boundary vertices belong to the boundary curves.
We report in Table 3, the errors and the convergence rates for the three methods. As in the previous situations, the
second-order boundary approach is doomed to a second-order of accuracy while the two other methods efficiently handle
the convection problem with curved boundaries with no oscillations and no artificial diffusion.
 R. Costa et al. / Applied Mathematical Modelling 54 (2018) 752–767 763

Table 3
Annulus problem for pure convection – Errors and convergence rates for the second-order, ROD1, and ROD2 methods with uniform triangular
Delaunay meshes.

Second-order method (Section 4.2.1)

P1 P3 P5

DOF E1 O1 E∞ O∞ E1 O1 E∞ O∞ E1 O1 E∞ O∞

736 1.01E−02 – 4.43E−02 – 6.78E−03 – 1.07E−02 – 6.88E−03 – 1.01E−02 –

2828 3.62E−03 1.52 2.35E−02 0.94 1.92E−03 1.88 2.92E−03 1.93 1.93E−03 1.89 2.85E−03 1.88
11,500 8.67E−04 2.04 1.07E−02 1.12 4.81E−04 1.97 7.19E−04 2.00 4.82E−04 1.98 7.18E−04 1.97
45,248 2.23E−04 1.98 1.91E−03 2.51 1.24E−04 1.98 1.86E−04 1.98 1.24E−04 1.98 1.86E−04 1.98
177,880 6.85E−05 1.72 4.07E−03 1.11 3.15E−05 2.00 4.72E−05 2.00 3.15E−05 2.00 4.72E−05 2.00

Reconstruction for Off-site Data (ROD1) method (Section 4.2.2)

P1 P3 P5

DOF E1 O1 E∞ O∞ E1 O1 E∞ O∞ E1 O1 E∞ O∞

736 8.75E−03 – 3.87E−02 – 2.35E−04 – 1.26E−03 – 3.48E−05 – 1.32E−04 –

2828 3.45E−03 1.38 2.15E−02 0.87 2.47E−05 3.35 1.09E−04 3.63 1.25E−06 4.95 5.26E−06 4.79
11,500 8.07E−04 2.07 1.03E−02 1.05 1.58E−06 3.91 9.50E−06 3.48 2.34E−08 5.67 1.14E−07 5.46
45,248 2.06E−04 2.00 1.81E−03 2.54 1.21E−07 3.76 8.04E−07 3.61 5.04E−10 5.60 3.11E−09 5.26
177,880 6.49E−05 1.69 4.05E−03 1.17 8.45E−09 3.89 9.60E−08 3.10 8.65E−12 5.94 5.76E−11 5.83

Reconstruction for Off-site Data (ROD2) method (Section 4.2.3)

P1 P3 P5

DOF E1 O1 E∞ O∞ E1 O1 E∞ O∞ E1 O1 E∞ O∞
736 8.75E−03 – 3.87E−02 – 2.36E−04 – 1.26E−03 – 3.57E−05 – 1.33E−04 –
2828 3.45E−03 1.38 2.15E−02 0.87 2.47E−05 3.35 1.10E−04 3.63 1.25E−06 4.98 5.27E−06 4.80
11,500 8.07E−04 2.07 1.03E−02 1.05 1.59E−06 3.91 9.50E−06 3.49 2.34E−08 5.68 1.14E−07 5.47
45,248 2.06E−04 2.00 1.81E−03 2.54 1.21E−07 3.76 8.04E−07 3.61 5.04E−10 5.60 3.11E−09 5.26
177,880 6.49E−05 1.69 4.05E−03 1.17 8.45E−09 3.89 9.61E−08 3.10 8.64E−12 5.94 5.76E−11 5.83

Fig. 8. Manufactured solution (left panel) and source term (right panel). (For interpretation of the colormap in this figure, the reader is referred to the
web version of this article).

6.2. Rose-shaped domain

We now consider a more complex geometry where the annulus is transformed by a diffeomorphism mapping which
consists in a periodic transformation of the boundaries in the following way:

   
x cos(θ ) x cos(θ )
I : 1 = RI (θ ; αI ) and E : 1 = RE (θ ; αE ) ,
x2 sin(θ ) x2 sin(θ )
764 R. Costa et al. / Applied Mathematical Modelling 54 (2018) 752–767

Fig. 9. Coarse uniform triangular Delaunay mesh (left panel) and uniform quadrilateral mesh (right panel) prescribed for the rose-shaped domain.

Table 4
Rose-shaped problem – Test 1 – Errors and convergence rates for the ROD2 method with uniform triangular Delaunay meshes.

P1 P3 P5

DOF E1 O1 E∞ O∞ E1 O1 E∞ O∞ E1 O1 E∞ O∞

667 2.21E−02 – 4.91E−02 – 3.59E−04 – 3.56E−03 – 2.20E−04 – 1.52E−03 –

2590 6.23E−03 1.86 1.40E−02 1.85 1.85E−05 4.38 2.30E−04 4.04 2.77E−06 6.45 3.04E−05 5.77
10,274 1.48E−03 2.09 4.49E−03 1.65 1.17E−06 4.00 2.02E−05 3.53 5.58E−08 5.67 1.12E−06 4.79
41,367 3.29E−04 2.16 1.21E−03 1.88 9.17E−08 3.66 1.53E−06 3.71 7.02E−10 6.28 2.07E−08 5.73
165,599 6.96E−05 2.24 3.45E−04 1.81 6.58E−09 3.80 1.07E−07 3.83 1.37E−11 5.68 3.85E−10 5.74

Table 5
Rose-shaped problem – Test 1 – Errors and convergence rates for the ROD2 method with uniform quadrilateral meshes.

P1 P3 P5

DOF E1 O1 E∞ O∞ E1 O1 E∞ O∞ E1 O1 E∞ O∞

660 6.42E−02 – 8.73E−02 – 1.41E−03 – 7.86E−03 – 9.75E−04 – 4.73E−03 –

2760 1.61E−02 1.93 2.23E−02 1.91 2.56E−04 2.39 8.45E−04 3.12 1.20E−05 6.15 8.33E−05 5.65
11,280 4.02E−03 1.97 5.60E−03 1.96 2.00E−05 3.62 8.48E−05 3.27 1.66E−07 6.08 1.08E−06 6.17
46,080 9.87E−04 2.00 1.38E−03 1.99 1.14E−06 4.07 5.93E−06 3.78 2.76E−09 5.83 2.00E−08 5.68
185,280 2.47E−04 1.99 3.45E−04 1.99 7.50E−08 3.91 3.84E−07 3.93 7.72E−11 5.14 3.42E−10 5.85

where (r, θ ) are the polar coordinates and RI (θ ; αI ) and RE (θ ; αE ), αI , αE ∈ R, are given by
 1
  1

RI (θ ; αI ) = rI 1 + sin(αI θ ) and RE (θ ; αE ) = rE 1 + sin(αE θ ) .
10 10
The global mapping from the Rose-shaped domain onto the annulus then reads
   R −r 
y1 x r − RI cos(θ )
→ 1 = T ( y1 , y2 ) = RI (θ ; αI ) + RE (θ ; αE )
E
.
y2 x2 RE − RI RE − RI sin(θ )
The manufactured solution on the Rose-shaped domain is then given by
ψ (x1 , x2 ) = φ (T −1 (x1 , x2 )).
Notice that we recover the annulus geometry with αI = αE = 0. The associated source term f is obtained from Eq. (1) while
homogeneous Dirichlet boundary condition still holds on the new boundaries I and E . All the simulations have been
carried out with κ = 1 and u = 1.

First test. The transformation is parametrized with αI = 3 and αE = 3 and we plot in Fig. 8 the manufactured solution and
the source term in the new domain.
 R. Costa et al. / Applied Mathematical Modelling 54 (2018) 752–767 765

Fig. 10. Manufactured solution (left panel) and source term (right panel). (For interpretation of the colormap in this figure, the reader is referred to the
web version of this article).

Fig. 11. Coarse uniform triangular Delaunay mesh (left panel) and uniform quadrilateral mesh (right panel) prescribed for the rose-shaped domain.

We carried out the simulations with successive refined regular triangular Delaunay meshes and also with quadrilateral
meshes, see Fig. 9, to show the ability of the method to handle different cell shapes.
We report in Table 4 the errors and the convergence rates obtained with the ROD2 method (Delaunay meshes) while
Table 5 reports the same informations for the quadrilateral meshes. We obtain the optimal convergence orders and no oscil-
lation is reported. Computations have also been carried out with the second-order boundary approximation (not presented
here) and we observe a second-order of convergence due to an inadequate treatment of the boundary condition.

Second test. The second test deals with a more wavy boundary setting αI = 3 and αE = 5. We plot in Fig. 10 the manufac-
tured solution and the source term.
As in the previous case, simulations with successive refined regular Delaunay meshes or quadrilateral meshes have been
carried out, see Fig. 11.
766 R. Costa et al. / Applied Mathematical Modelling 54 (2018) 752–767

Table 6
Rose-shaped problem – Test 2 – Errors and convergence rates for the ROD2 method with uniform triangular Delaunay meshes.

P1 P3 P5

DOF E1 O1 E∞ O∞ E1 O1 E∞ O∞ E1 O1 E∞ O∞

645 2.71E−02 – 6.97E−02 – 6.80E−04 – 8.76E−03 – 5.12E−04 – 5.86E−03 –

2550 7.56E−03 1.86 2.30E−02 1.61 4.49E−05 3.95 7.62E−04 3.55 1.29E−05 5.36 2.45E−04 4.62
10,244 1.89E−03 2.00 7.30E−03 1.65 3.71E−06 3.59 1.28E−04 2.57 2.98E−07 5.42 1.39E−05 4.12
40,789 4.10E−04 2.21 1.97E−03 1.89 2.53E−07 3.89 6.70E−06 4.27 9.53E−09 4.98 4.91E−07 4.84
162,011 9.63E−05 2.10 5.92E−04 1.75 1.89E−08 3.76 9.37E−07 2.85 1.20E−10 6.34 1.07E−08 5.55

Table 7
Rose-shaped problem – Test 2 – Errors and convergence rates for the ROD2 method with uniform quadrilateral meshes.

P1 P3 P5

DOF E1 O1 E∞ O∞ E1 O1 E∞ O∞ E1 O1 E∞ O∞

660 6.91E−02 – 1.11E−01 – 5.06E−03 – 3.27E−02 – 5.37E−03 – 3.60E−02 –

2760 1.74E−02 1.93 2.85E−02 1.90 3.94E−04 3.57 3.57E−03 3.10 1.33E−04 5.17 1.61E−03 4.35
11,280 4.37E−03 1.96 7.13E−03 1.97 4.70E−05 3.02 3.83E−04 3.17 2.27E−06 5.78 5.07E−05 4.91
46,080 1.08E−03 1.99 1.77E−03 1.98 3.36E−06 3.75 2.50E−05 3.88 4.85E−08 5.47 1.07E−06 5.48
185,280 2.71E−04 1.99 4.42E−04 1.99 2.05E−07 4.02 1.57E−06 3.98 1.11E−09 5.42 1.41E−08 6.22

We report in Tables 6 and 7 the errors and the convergence rates obtained with the ROD2 method and we confirm
the ability of the scheme to preserve the optimal order in function of the polynomial degree used for the reconstruction
procedure.

7. Conclusions

We have presented a very high-order finite volume scheme to solve the steady-state bi-dimensional convection–diffusion
problem based on a new class of polynomial reconstructions. Two approaches were proposed to overcome the second-order
accuracy limitation when dealing with non-polygonal domain and Dirichlet boundary conditions. The first one consists in
analytically constraining the boundary element reconstructions in order to satisfy the boundary condition at a point on
the physical domain boundary. Such approach differs from the Olliver-Gooch and Van Altena approach in the sense that
the flux calculation is performed on the straight edge and no curved element is necessary. The proposed ROD method
consists in constraining the boundary reconstructions by a posteriori computing the associated free parameter such that the
reconstructions satisfy appropriately the boundary condition. This procedure relies on the fact that the least-squares matrix
associated to the reconstruction is decoupled from the boundary parameterization and, therefore, is less sensitive to the
boundary location.
Several numerical tests considering simple and complex curved domains were simulated to observe that we achieve
effective optimal order of accuracy both for structured and unstructured meshes for the two-dimensional linear steady-state
convection–diffusion problem. A pure convection problem (hyperbolic scalar equation) was also tested and optimal order
accuracy rates were achieved without any reported oscillations.
This work represents a proof of concept showing that very high-order of accuracy Finite Volume scheme on unstructured
can handle curved boundary conditions at the optimal order of accuracy without the need for a boundary fitted mesh or
complex transformations. For future works we plan to extend this approach to other boundary conditions (Neumann, Robin),
to unsteady systems (Euler, Navier–Stokes) with time evolving domain (piston, pulsating interfaces, etc.). We also plan to
investigate the extension of ROD to unstructured 3D mesh. Even if, conceptually speaking, the ROD method does not depend
on the space dimension, the machinery needed in three-dimensional will demand delicate validation and verification.

Acknowledgments

This research was financed by FEDER Funds through Programa Operational Fatores de Competitividade – COMPETE and
by Portuguese Funds FCT – Fundação para a Ciência e a Tecnologia, within the Strategic Project UID/MAT/0 0 013/2013. R. L.
and R. C. thank the financial support of the “International Centre for Mathematics and Computer Science in Toulouse” (CIMI)
partially supported by ANR-11-LABX-0040-CIMI within the program ANR-11-IDEX-0 0 02-02. The authors would also like to
thank the referee for the additional informations and references.

References

[1] F. Bassi, S. Rebay, High-order accurate discontinuous finite element solution of the 2d Euler equations, J. Comput. Phys. 138 (2) (1997) 251–285.
[2] L. Krivodonova, M. Berger, High-order accurate implementation of solid wall boundary conditions in curved geometries, J. Comput. Phys. 211 (2006)
492–512.
[3] Z.J. Wang, High-order computational fluid dynamics tools for aircraft design, Philos. Trans. R. Soc. A 372 (2014) 20130318.
 R. Costa et al. / Applied Mathematical Modelling 54 (2018) 752–767 767

[4] Z.J. Wang, High-order methods for the euler and Navier–Stokes equations on unstructured grids, Prog. Aerosp. Sci. 43 (1–3) (2007) 1–41.
[5] D. Mavriplis, Results from the 3rd Drag Prediction Workshop Using the NSU3d Unstructured Mesh Solver, AIAA Paper 2007-256 (2007).
[6] I. Ergatoudis, B.M. Irons, O.C. Zienkiewicz, Curved, isoparametric, quadrilateral elements for finite element analysis, Int. J. Solids Struct. 4 (1968) 31–42.
[7] P.G. Ciarlet, P.A. Raviart, Interpolation over curved elements with applications to finite element methods, Comput. Methods Appl. Mech. Eng. 1 (1972)
217–249.
[8] M. Zlámal, Curved elements in the finite element method i, SIAM J. Numer. Anal. 10 (1973) 229–240.
[9] M. Zlámal, Curved elements in the finite element method II, SIAM J. Numer. Anal. 11 (1974) 347–368.
[10] B. Cockburn, G.E. Karniadakis, C.W. Shu, The development of discontinuous Galerkin methods, in: B. Cockburn, G.E. Karniadakis, C.-W. Shu (Eds.),
Discontinuous Galerkin Methods. Theory, Computation and Applications, 11, Springer Verlag, Berlin, Heidelberg, 20 0 0, pp. 5–50. Lecture Notes in
Computational Sicence and Engineering.
[11] F. Hindenlang, T. Bolemann, C.D. Munz, Mesh curving techniques for high order discontinuous Galerkin simulations, in: N. Kroll, C. Hirsch, F. Bassi,
C. Johnston, K. Hillewaert (Eds.), Industrialization of High-Order Methods – A Top-Down Approach, 128, Springer International Publishing, 2015,
pp. 133–152.
[12] Z.J. Wang, Y. Liu, Extension of the spectral volume method to high-order boundary representation, J. Comput. Phys. 211 (2006) 154–178.
[13] W.J. Gordon, C.A. Hall, Construction of curvilinear co-ordinate systems and applications to mesh generation, Int. J. Numer. Methods Eng. 7 (4) (1973)
461–477.
[14] W.J. Gordon, C.A. Hall, Transfinite element methods: blending-function interpolation over arbitrary curved element domains, Numer. Math. 21 (2)
(1973) 109–129.
[15] C. Geuzaine, J.F. Remacle, Gmsh: a three-dimensional finite element mesh generator with built-in pre- and post-processing facilities, Int. J. Numer.
Methods Eng. 79 (2009) 1309–1331.
[16] N. Balakrishnan, G. Fernandez, Wall boundary conditions for inviscid compressible flows on unstructured meshes, Int. J. Numer. Methods Fluids 28
(1998) 1481–1501.
[17] B. Cockburn, M. Solano, Solving dirichlet boundary-value problems on curved domains by extensions from subdomains, SIAM J. Sci. Comput. 34 (2012)
497–519.
[18] B. Cockburn, M. Solano, Solving convection–diffusion problems on curved domains by extensions from subdomains, J. Sci. Comput. 59 (2014) 512–543.
[19] W. Qiu, M. Solano, P. Vega, A high order HDG method for curved-interface problems via approximations from straight triangulations, J. Sci. Comput.
69 (2016) 1384–1407.
[20] E. Burman, S. Claus, P. Hansbo, M.G. Larson, A. Massing, CutFEM: discretizing geometry and partial differential equations, Int. J. Numer. Methods Eng.
104 (2015) 472–501.
[21] C. Lehrenfeld, A. Reusken, Analysis of a high order unfitted finite element method for an elliptic interface problem, IMA J. Numer. Anal. (2017),
doi:10.1093/imanum/drx041. (accessed 28.09.2017).
[22] C. Lehrenfeld, High order unfitted finite element methods on level set domains using isoparametric mappings, Comput. Methods Appl. Mech. Eng. 300
(2016) 716–733.
[23] C. Ollivier-Gooch, M.V. Altena, A high-order-accurate unstructured mesh finite-volume scheme for the advection-diffusion equation, J. Comput. Phys.
181 (2) (2002) 729–752.
[24] C. Michalak, C. Ollivier-Gooch, Unstructured High-order Accurate Finite Volume Solutions of the Navier–Stokes Equations, AIAA Paper 2009-954 (2009).
[25] C. Ollivier-Gooch, A. Nejat, C. Michalak, On Obtaining High-order Finite-volume Solutions to the Euler Equations on Unstructured Meshes, AIAA Paper
20 07-4464 (20 07).
[26] C. Ollivier-Gooch, A. Nejat, C. Michalak, On obtaining and verifying high-order unstructured finite volume solutions to the euler equations, AIAA J. 47
(9) (2009) 2105–2120.
[27] A. Boularas, S. Clain, F. Baudoin, A sixth-order finite volume method for diffusion problem with curved boundaries, Appl. Math. Model. 41 (2017)
401–422.
[28] T.J. Barth, P.O. Frederickson, Higher Order Solution of the Euler Equations on Unstructured Grids Using Quadratic Reconstruction, AIAA Paper 90-0013
(1990).
[29] T.J. Barth, Recent Developments in High Order k-exact Reconstruction on Unstructured Meshes, AIAA Paper 93-0668 (1993).
[30] S. Clain, G.J. Machado, J.M. Nóbrega, R.M.S. Pereira, A sixth-order finite volume method for the convection–diffusion problem with discontinuous
coefficients, Comput. Methods Appl. Mech. Eng. 267 (2013) 43–64.
[31] A. Ern, J.L. Guermond, Theory and Practice of Finite Elements, 159, Springer Verlag, New-York, 2004.